cntk.random package

Functions that generate random numbers with respect to designated distributions.

bernoulli(shape, dtype=default_override_or(np.float32), mean=0.5, seed=auto_select, name='')[source]

Generates samples from the Bernoulli distribution with success probability mean.

Parameters:
  • shape (tuple) – shape of the output (entries are independent random draws)
  • dtype (np.float32 or np.float64 or np.float16) – data type. Default is np.float32.
  • mean (float) – success probability
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> b = C.random.bernoulli((2,3), seed=98052)
>>> b.eval(device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
array([[ 1.,  1.,  0.],
       [ 1.,  0.,  0.]], dtype=float32)
bernoulli_like(x, mean=0.5, seed=auto_select, name='')[source]

Generates samples from the Bernoulli distribution with success probability mean.

Parameters:
  • x – cntk variable (input, output, parameter, or constant) from which to copy the shape, data type, and dynamic axes.
  • mean (float) – success probability
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> p = C.placeholder()
>>> bp = C.random.bernoulli_like(p, seed=98052)
>>> x = C.sequence.input_variable(1)
>>> bx = bp.replace_placeholders({p:x})
>>> x0 = np.zeros((1,3,1), dtype=np.float32)
>>> bx.eval({x:x0}, device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
[array([[ 1.],
       [ 1.],
       [ 0.]], dtype=float32)]
gumbel(shape, dtype=default_override_or(np.float32), loc=0.0, scale=1.0, seed=auto_select, name='')[source]

Generates samples from the Gumbel distribution with location loc and scale scale.

Parameters:
  • shape (tuple) – shape of the output (entries are independent random draws)
  • dtype (np.float32 or np.float64 or np.float16) – data type. Default is np.float32.
  • loc (float) – location of the distribution
  • scale (float) – scale of the distribution
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> g = C.random.gumbel((2,3), seed=98052)
>>> g.eval(device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
array([[-0.987713, -0.522298,  0.425918],
       [-1.019599,  5.435177,  1.586071]], dtype=float32)

See also

The Gumbel-Max Trick.

gumbel_like(x, mean=0.0, scale=1.0, seed=auto_select, name='')[source]

Generates samples from the Gumbel distribution with location loc and scale scale.

Parameters:
  • x – cntk variable (input, output, parameter, or constant) from which to copy the shape, data type, and dynamic axes.
  • loc (float) – location of the distribution
  • scale (float) – scale of the distribution
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> x = C.constant(np.zeros((2,3,4), dtype=np.float32))
>>> g = C.random.gumbel_like(x, seed=98052)
>>> s = g.eval(device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
>>> np.round(s, 4)
array([[[-0.9877, -0.5223,  0.4259, -1.0196],
        [ 5.4352,  1.5861,  5.0608,  2.0668],
        [-0.2135,  1.0139,  3.1217, -1.4834]],

       [[ 0.4507,  0.6325,  2.1682,  0.4463],
        [-0.6583,  0.1147, -0.3144, -0.7925],
        [ 1.9773, -0.3627, -0.4566, -0.2368]]], dtype=float32)

See also

The Gumbel-Max Trick.

normal(shape, dtype=default_override_or(np.float32), mean=0.0, scale=1.0, seed=auto_select, name='')[source]

Generates samples from the normal distribution with mean mean and standard deviation scale.

Parameters:
  • shape (tuple) – shape of the output (entries are independent random draws)
  • dtype (np.float32 or np.float64 or np.float16) – data type. Default is np.float32.
  • mean (float) – mean of the distribution
  • scale (float) – scale (standard deviation) of the distribution
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> z = C.random.normal((2,3), seed=98052)
>>> z.eval(device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
array([[ 1.803254,  0.995395, -0.631974],
       [-1.736721,  0.005615, -0.340025]], dtype=float32)
normal_like(x, mean=0.0, scale=1.0, seed=auto_select, name='')[source]

Generates samples from the normal distribution with mean mean and standard deviation scale.

Parameters:
  • x – cntk variable (input, output, parameter, or constant) from which to copy the shape, data type, and dynamic axes.
  • mean (float) – mean of the distribution
  • scale (float) – scale (standard deviation) of the distribution
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> x = C.parameter((2,3,4))
>>> z = C.random.normal_like(x, seed=98052)
>>> z.eval(device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
array([[[ 1.803254,  0.995395, -0.631974, -1.736721],
        [ 0.005615, -0.340025, -0.011913, -0.236371],
        [-1.207685, -0.495846,  0.037022, -1.220596]],

       [[ 0.872981,  0.654405, -0.111421, -0.544074],
        [ 1.543746, -0.63555 , -1.072869, -0.379701],
        [ 0.592069, -1.035192,  1.679303, -0.391963]]], dtype=float32)
uniform(shape, dtype=default_override_or(np.float32), low=0.0, high=1.0, seed=auto_select, name='')[source]

Generates samples from the uniform distribution in the interval [low,`high`).

Parameters:
  • shape (tuple) – shape of the output (entries are independent random draws)
  • dtype (np.float32 or np.float64 or np.float16) – data type. Default is np.float32.
  • low (float) – lower end of the range of the random numbers
  • high (float) – upper end of the range of the random numbers
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> u = C.random.uniform((2,3), seed=98052)
>>> u.eval(device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
array([[ 0.931785,  0.814722,  0.479606],
       [ 0.937468,  0.004351,  0.185131]], dtype=float32)
uniform_like(x, low=0.0, high=1.0, seed=auto_select, name='')[source]

Generates samples from the uniform distribution in the interval [low,`high`).

Parameters:
  • x – cntk variable (input, output, parameter, or constant) from which to copy the shape, data type, and dynamic axes.
  • low (float) – lower end of the range of the random numbers
  • high (float) – upper end of the range of the random numbers
  • seed (int) – pseudo random number generator seed (default: automatically select a unique seed)
  • name (str, optional) – the name of the Function instance in the network
Returns:

Function

Examples

>>> x = C.input_variable(4)
>>> x0 = np.zeros((3,4), dtype=np.float32)
>>> u = C.random.uniform_like(x, seed=98052)
>>> u.eval({x:x0}, device=C.cpu()) # explicitly setting cpu because this is tested on multiple platforms; leave it unspecified in your code
array([[ 0.931785,  0.814722,  0.479606,  0.937468],
       [ 0.004351,  0.185131,  0.00632 ,  0.118901],
       [ 0.710054,  0.304273,  0.043126,  0.987818]], dtype=float32)